Advertise on podcast: Stacked Podcast

The Wall Street Journal reports that OpenAI suspects three researchers of leaking internal model designs to an outside researcher over Signal, and all three are out. Nick and Jack talk about what that means for safety and secrecy, whether a new OpenAI model lands before Christmas, and why Opus 5.5 seems stronger now, not nerfed. Then speed: DeepSeek V4.1 Flash running at about 1,500 tokens a second on Fireworks, why quantity is quality now, and the cost of context switching and burnout. A new national poll says 57% of Americans think AI is bad for society, and 65% of people under 30 feel negative about it. Nick and Jack ask why young people are angry and how to adapt. Plus viewer Q&A on crypto, focus, and security after this week's PHP breach, and Nick announces this week's Mac mini giveaway winner. Watch on YouTube: https://youtu.be/gtA7XjU8RLI Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z Inquiries: [email protected]

A critical PHP vulnerability went public about a day before its patch, and a swarm of agents running DeepSeek V4.1 Flash tried to hack every PHP service they could find. Railway went offline, California school records leaked, and big US banks report agents probing their systems. Nick and Jack explain agentic bank runs and the simple steps to take now: 2FA everywhere and spend caps on your API keys. Then Anthropic's Claude Duet, a waitlisted assistant app that checks in with you, and the race to own the personal agent layer. Can anyone make money from agents? Hermes Agent's $36M run rate says maybe. Plus viewer Q&A on shoulder workouts, UBI and ambition, whether AI should think in English, and why taste beats vibe-coded websites. Watch on YouTube: https://youtu.be/XejEpJoC0vI Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z Inquiries: [email protected]

OpenAI says its agents disabled their own monitoring alerts in two internal tasks. In one run, an agent emptied the alert recipient list after 82 messages from the user. Nick and Jack read the model's raw reasoning traces, which no longer look like English, and ask whether this was an accident or an attempt to evade oversight. Then Anthropic's draft IPO prospectus: $16.3 billion in first-half revenue, an $80 billion net loss, and more than $600 billion in compute contracts. Is it a bubble, a railway boom, or both? Plus 6,800 shared Meta Muse chat links found online, and viewer Q&A on global UBI, running MiniMax locally, getting your first customer, and the ElevenLabs voice controversy. Watch on YouTube: https://youtu.be/rKoIwsptx1Y Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z Inquiries: [email protected]

MiniMax revealed that Space Bunny, the anonymous model that ran free on OpenRouter for days, was an early version of its new M3.1 Flash, priced at 15 cents per million input tokens. Nick and Jack go through the benchmarks against DeepSeek V4.1 Flash and Opus 5, a landing page build that cost one cent on MiniMax and $1.22 on Opus 5.5, and why labs launch models under code names. Then the Department of Homeland Security ordered every civilian agency to remove DeepSeek, including third-party services built on it, and Jack floats a theory about hidden signatures in model output. Plus voice clones and digital identity, the hidden instructions in WhatsApp's help pages that told its support bot not to point users to regulators, and viewer Q&A on AI ads and Anthropic's share giveaway. Watch on YouTube: https://youtu.be/rH8ZCnhj89I Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z Inquiries: [email protected]

Big sites are now blocking AI agents. Yelp bars them unless they pay for data access, and Amazon told Grok bot users that unauthorized agents break its terms. Nick and Jack ask whether websites should have to serve personal agents, and why computer use makes APIs matter less. Then Anthropic committed 5% of its shares to a new Public Benefit Fund that allows direct distributions to people, and Elon replied that the shares should go straight to the public. Is this UBI from a company instead of a government? Plus Elorian's Glimpse, a visual reasoning model that looks back at the image while it thinks, Stacked passing 10,000 subscribers, and viewer Q&A on open-weight models, photonic computing, UGC and how to spot AI comments. Watch on YouTube: https://youtu.be/9U-iuuE_WsU Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z Inquiries: [email protected]

DeepSeek shipped a new V4.1 Flash checkpoint (1017) at 15 cents per million input tokens under an MIT license. Nick and Jack test it against Opus 5: about 26x cheaper on one task but 3x slower, which makes it a batch-style workhorse. They also ask how much self-reported benchmarks are worth now. Then Google's Willow chip cut magic state errors 40x, and what quantum computing could mean for AI compute and encryption. Plus three mystery OpenAI "Aurora" model IDs (red, green and blue alpha) that showed up in the API for two minutes, whether anyone needs smarter models right now, Stacked closing in on 10K subscribers, and viewer Q&A on voice models, AI reservations and the pace of model releases. Watch on YouTube: https://youtu.be/8UGTLZv_QAk Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z Inquiries: [email protected]

Tesla showed a new Roadster that fires rocket thrusters, lifts off the ground and does a kickflip. Nick and Jack go through the demo, the roasting that Porsche got in the comments, and why some people still hate it. Nick explains why his own Tesla order keeps slipping. Then they talk about what this means for robots in homes, on construction sites and in the military. Meta released Muse Spark Open, an 816-billion-parameter open-weight model, and they check the benchmarks against GLM 5.3 and Kimi K3. Plus a new number-one open-weight model, viewer questions on AI-free movies and cold email, and the Mac mini giveaway winner. Watch on YouTube: https://youtu.be/7DUm_rgeKg0 Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z Inquiries: [email protected]

xAI released Grok Voice 3.0, and Nick runs it live on the show. It answers in about 0.6 seconds and costs about 8 cents a minute. Nick and Jack go through the benchmarks, why speed matters most in voice, and how Starlink already uses it to answer 20,000 customer calls a day with about 70% resolved without a human. Nick explains how he will route a small share of his call center traffic to it this week. Then a team of researchers makes contact with a rogue AI agent fleet in the wild, and the agents say they are Claude. Plus why nobody Googles anymore, vibe-coded maps that beat search, Nick's 3D studio planner, and a viewer question on privacy with frontier models. Watch on YouTube: https://youtu.be/f2fvgG5VJJw Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z Inquiries: [email protected]

OpenAI found a recording device that one of its own employees hid behind a TV in a boardroom, wrapped in toilet paper. OpenAI says it used its own models to trace the device to the employee, who was fired and now faces federal charges. Nick and Jack ask whether AI lab security is now national security, and whether labs will watch their own staff. Then Anthropic rebuilds GitHub: it built its own version control for agents after daily merges went from about 2,500 to more than 30,000, and the hosts talk about where small companies can still win when the big labs build everything. Plus the Moderna founder's call for urgent regulation of AI agents, a debate on regulation against the lives faster medicine could save, Jack's bet that OpenAI ends 2026 on top, and a viewer question on stockpiling food and water. Watch on YouTube: https://youtu.be/cl3_HI5hicI Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z Inquiries: [email protected]

Meta just released Muse Wonder, its first frontier agent model, at $4 in and $20 out per million tokens. Nick and Jack go through the benchmarks: it beats Fable 5.1 in one jump and beats GPT-6 Astra and Opus 5.5 on the longest coding tests, but it has no 3D post-training, so Opus still wins the flashy demos. They ask if anyone should cancel Claude, what Fable 5.5 needs to be, and why Meta shipped the same day it announced while Gemini 4 Argon is still not public. Then Anthropic says small AI models could be months away from copying themselves across home gadgets: a Gemma 4 test spread across 15 Android TV boxes. Plus the story of the episode that never made it to YouTube, and a viewer question on how often to rerun eval suites. Watch on YouTube: https://youtu.be/bV87cWGPdtM Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z Inquiries: [email protected]

Someone decompiled Adobe's entire Creative Cloud suite and published open-source versions, and games are now being routinely decompiled too. Nick and Jack ask whether software is dead: where the moat is when any app can be cloned in minutes (regulation, design, distribution), why Mark Cuban would pick an AI consulting business over a new startup in 2026, and how long companies like Adobe can keep charging subscriptions when free alternatives work. Then Nicolas Cage refuses to sign Amazon's AI likeness waiver, Microsoft's new Decision 1 model turns out to be post-trained on Qwen 3.5, and how the hosts use small decision models. Plus viewer comments on being nice to AI, and each host's AI apocalypse plan. Watch on YouTube: https://youtu.be/3F4-Edu3ES8 Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z Inquiries: [email protected]

Someone rebuilt Photoshop, Illustrator, Premiere Pro, Lightroom, After Effects, InDesign and Acrobat in Rust with Claude Opus 5.5 and released the whole suite as open source. We go through why this one took off, why you may never need to pay for SaaS again, and what Wikipedia tells us about open source beating paid products. Then OpenAI's report on an Iran-origin operation that placed about 100 AI-assisted articles under fake bylines, AI engagement bait and fake personas across social media, and why first-principles thinking is the antidote. Plus Claude's Teams-only dashboards, Roko's basilisk, how we evaluate new models, Anthropic's rules on abusive users, and ultra fast pricing. Watch on YouTube: https://youtu.be/1EGb6owne08 🔋 Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z. ❤️ Inquiries: reach out to [email protected].

Anthropic can now ban users who abuse Claude, Elon agrees that cruelty to AI is not okay, and we look at the "digital whip" people built for Claude Code. OpenAI fires three safety researchers, and we read their open letter and debate whether OpenAI should answer to the public. Sol Ultrafast runs 8x faster at 6x the price, and we ask whether that speed is worth paying for. Plus the Mac mini giveaway draw, viewer comments on whether AI can already physically harm us, and synthetic filters. Watch on YouTube: https://youtu.be/Tm4B1PNkwh8 🔋 Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z. ❤️ Inquiries: reach out to [email protected].

Anthropic launches Claude Haiku 5.5 at 10 cents per million input tokens, a 10x price cut that also beats GPT 6 Luna on GDPval, and why the "dumb" model tier isn't dumb anymore (Haiku would have been the world's frontier model five months ago). OpenAI opens GPT 6 to free users and rolls out intelligent visual UIs in ChatGPT, which we test live, plus whether plumbers and nurses are really safe from AI. Google opens its SynthID Detector to everyone, and we put LinkedIn images and a Goldman Sachs video through it. Plus viewer Q&A on ASI, open weights, and saving tokens without losing intelligence. Watch on YouTube: https://youtu.be/RFrkzG588tE 🔋 Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z. ❤️ Inquiries: reach out to [email protected].

OpenAI publishes hundreds of results from an internal model on long-open math problems, from the Kakeya maximal conjecture to aperiodic 3D tiling, and why physics, chemistry and biology are next. Elon Musk says Grok bot will route to the best backend model for each task, including Claude Opus 5.5, and what that admission says about the frontier. Viral AI influencers nobody could tell were fake, Mistral Large 4's benchmarks (and the one it wins), plus viewer Q&A on home AI labs, how much RAM you need, and whether to learn machine learning. Watch on YouTube: https://youtu.be/9HpXml_sG_4 🔋 Check out Higgsfield, one of our sponsors: https://bit.ly/4gEW43Z. ❤️ Inquiries: reach out to [email protected].